The LLVM Compiler Framework and Infrastructure

Transcription

1 The LLVM Compiler Framework and Infrastructure Program Analysis Original Slides by David Koes (CMU) Substantial portions courtesy Chris Lattner and Vikram Adve

2 LLVM Compiler System The LLVM Compiler Infrastructure Provides reusable components for building compilers Reduces the time/cost to build a new compiler Build static compilers, JITs, trace-based optimizers,... The LLVM Compiler Framework End-to-end compilers using the LLVM infrastructure C and C++ gcc frontend Backends for C, X86, Sparc, PowerPC, Alpha, Arm, Thumb, IA-64 2

3 Three primary LLVM components The LLVM Virtual Instruction Set The common language- and target-independent IR Internal (IR) and external (persistent) representation A collection of well-integrated libraries Analyses, optimizations, code generators, JIT compiler, garbage collection support, profiling, A collection of tools built from the libraries Assemblers, automatic debugger, linker, code generator, compiler driver, modular optimizer, 3

4 Tutorial Overview Introduction to the running example LLVM C/C++ Compiler Overview High-level view of an example LLVM compiler The LLVM Virtual Instruction Set IR overview and type-system LLVM C++ IR and important API s Basics, PassManager, dataflow, ArgPromotion Alias Analysis in LLVM 4

5 Running example: arg promotion Consider use of by-reference parameters: int callee(const int &X) { return X+1; int caller() { return callee(4); We want: int callee(int X) { return X+1; int caller() { return callee(4); compiles to int callee(const int *X) { return *X+1; // memory load int caller() { int tmp; // stack object tmp = 4; // memory store return callee(&tmp); Eliminated load in callee Eliminated store in caller Eliminated stack slot for tmp 5

6 Why is this hard? Requires interprocedural analysis: Must change the prototype of the callee Must update all call sites we must know all callers What about callers outside the translation unit? Requires alias analysis: Reference could alias other pointers in callee Must know that loaded value doesn t change from function entry to the load Must know the pointer is not being stored through Reference might not be to a stack object! 6

7 Tutorial Overview Introduction to the running example LLVM C/C++ Compiler Overview High-level view of an example LLVM compiler The LLVM Virtual Instruction Set IR overview and type-system LLVM C++ IR and important API s Basics, PassManager, dataflow, ArgPromotion Alias Analysis in LLVM 7

8 The LLVM C/C++ Compiler From the high level, it is a standard compiler: Compatible with standard makefiles Uses GCC 4.2 C and C++ parser Generates native executables/object files/assembly Distinguishing features: Uses LLVM optimizers, not GCC optimizers Pass -emit-llvm to output LLVM IR -S: human readable assembly -c: efficient bitcode binary 8

9 Looking into events at compile-time C/C++ file llvm-gcc/llvm-g++ -O -S assembly IR GENERIC GIMPLE (tree-ssa) LLVM IR Machine Code IR -emit-llvm LLVM asm >50 LLVM Analysis & Optimization Passes: Dead Global Elimination, IP Constant Propagation, Dead Argument Elimination, Inlining, Reassociation, LICM, Loop Opts, Memory Promotion, Dead Store Elimination, ADCE, 9

10 Looking into events at link-time LLVM bitcode.o file LLVM bitcode.o file llvm-ld executable LLVM Linker >30 LLVM Analysis & Optimization Passes Optionally internalizes : marks most functions as internal, to improve IPO Link-time Optimizer -native.bc file for LLVM JIT Native executable Perfect place for argument promotion optimization! 10

11 Goals of the compiler design Analyze and optimize as early as possible: Compile-time opts reduce modify-rebuild-execute cycle Compile-time optimizations reduce work at link-time (by shrinking the program) All IPA/IPO make an open-world assumption Thus, they all work on libraries and at compile-time Internalize pass enables whole program optzn One IR (without lowering) for analysis & optzn Compile-time optzns can be run at link-time too! The same IR is used as input to the JIT IR design is the key to these goals! 11

12 Tutorial Overview Introduction to the running example LLVM C/C++ Compiler Overview High-level view of an example LLVM compiler The LLVM Virtual Instruction Set IR overview and type-system LLVM C++ IR and important API s Basics, PassManager, dataflow, ArgPromotion Alias Analysis in LLVM 12

13 Goals of LLVM IR Easy to produce, understand, and define! Language- and Target-Independent AST-level IR (e.g. ANDF, UNCOL) is not very feasible Every analysis/xform must know about all languages One IR for analysis and optimization IR must be able to support aggressive IPO, loop opts, scalar opts, high- and low-level optimization! Optimize as much as early as possible Can t postpone everything until link or runtime No lowering in the IR! 13

14 LLVM Instruction Set Overview #1 Low-level and target-independent semantics RISC-like three address code Infinite virtual register set in SSA form Simple, low-level control flow constructs Load/store instructions with typed-pointers IR has text, binary, and in-memory forms bb: ; preds = %bb, %entry %i.1 = phi i32 [ 0, %entry ], [ %i.2, %bb ] %AiAddr = getelementptr float* %A, i32 %i.1 for (i = 0; i < N; call float* %AiAddr, %pair* %P ) %i.2 = add i32 %i.1, 1 ++i) %exitcond = icmp eq i32 %i.2, %N Sum(&A[i], &P); br i1 %exitcond, label %return, label %bb 14

15 LLVM Instruction Set Overview #2 High-level information exposed in the code Explicit dataflow through SSA form Explicit control-flow graph (even for exceptions) Explicit language-independent type-information Explicit typed pointer arithmetic Preserve array subscript and structure indexing bb: ; preds = %bb, %entry %i.1 = phi i32 [ 0, %entry ], [ %i.2, %bb ] %AiAddr = getelementptr float* %A, i32 %i.1 for (i = 0; i < N; call float* %AiAddr, %pair* %P ) %i.2 = add i32 %i.1, 1 ++i) %exitcond = icmp eq i32 %i.2, %N Sum(&A[i], &P); br i1 %exitcond, label %return, label %bb 15

16 LLVM Type System Details The entire type system consists of: Primitives: integer, floating point, label, void no signed integer types arbitrary bitwidth integers (i32, i64, i1) Derived: pointer, array, structure, function, vector, No high-level types: type-system is language neutral! Type system allows arbitrary casts: Allows expressing weakly-typed languages, like C Front-ends can implement safe languages Also easy to define a type-safe subset of LLVM See also: docs/langref.html 16

17 Lowering source-level types to LLVM Source language types are lowered: Rich type systems expanded to simple type system Implicit & abstract types are made explicit & concrete Examples of lowering: References turn into pointers: T& T* Complex numbers: complex float { float, float Bitfields: struct X { int Y:4; int Z:2; { i32 Inheritance: class T : S { int X; { S, i32 Methods: class T { void foo(); void foo(t*) Same idea as lowering to machine code 17

18 LLVM Program Structure Module contains Functions/GlobalVariables Module is unit of compilation/analysis/optimization Function contains BasicBlocks/Arguments Functions roughly correspond to functions in C BasicBlock contains list of instructions Each block ends in a control flow instruction Instruction is opcode + vector of operands All operands have types Instruction result is typed 18

19 Our example, compiled to LLVM int callee(const int *X) { return *X+1; // load int caller() { int T; // on stack T = 4; // store return callee(&t); All loads/stores are Stack allocation is explicit in the LLVM explicit in LLVM representation define internal %X) { entry: %tmp2 = load i32* %X %tmp3 = add i32 %tmp2, 1 ret i32 %tmp3 define internal { entry: %T = alloca i32 store i32 4, i32* %T %tmp1 = call i32* %T ) ret i32 %tmp1 19

20 Our example, desired transformation define %X) { %tmp2 = load i32* %X %tmp3 = add i32 %tmp2, 1 ret i32 %tmp3 define { %T = alloca i32 store i32 4, i32* %T %tmp1 = call i32* %T ) ret i32 %tmp1 define internal %X.val) { %tmp3 = add i32 %X.val, 1 ret i32 %tmp3 define internal { %T = alloca i32 store i32 4, i32* %T %Tval = load i32* %T %tmp1 = call i32 %Tval ) ret i32 %tmp1 Other transformation Insert Change Update load all the instructions call prototype sites of (-mem2reg) cleans up for into the callee function callers the rest define internal { %tmp1 = call i32 4 ) ret i32 %tmp1 20

21 Tutorial Overview Introduction to the running example LLVM C/C++ Compiler Overview High-level view of an example LLVM compiler The LLVM Virtual Instruction Set IR overview and type-system LLVM C++ IR and important API s Basics, PassManager, dataflow, ArgPromotion Alias Analysis in LLVM 21

22 LLVM Coding Basics Written in modern C++, uses the STL: Particularly the vector, set, and map classes LLVM IR is almost all doubly-linked lists: Module contains lists of Functions & GlobalVariables Function contains lists of BasicBlocks & Arguments BasicBlock contains list of Instructions Linked lists are traversed with iterators: Function *M = for (Function::iterator I = M->begin(); I!= M->end(); ++I) { BasicBlock &BB = *I;... See also: docs/programmersmanual.html 22

23 LLVM Coding Basics cont. BasicBlock doesn t provide a reverse iterator Highly obnoxious when doing the assignment for(basicblock::iterator I = bb->end(); I!= bb->begin(); ) { --I; Instruction *insn = I; Traversing successors of a BasicBlock: for (succ_iterator SI = succ_begin(bb), E = succ_end(bb); SI!= E; ++SI) { BasicBlock *Succ = *SI; C++ is not Java primitive class variable not automatically initialized you must manage memory virtual vs. non-virtual functions and much much more 23

24 LLVM Pass Manager Compiler is organized as a series of passes : Each pass is one analysis or transformation Types of Pass: ModulePass: general interprocedural pass CallGraphSCCPass: bottom-up on the call graph FunctionPass: process a function at a time LoopPass: process a natural loop at a time BasicBlockPass: process a basic block at a time Constraints imposed (e.g. FunctionPass): FunctionPass can only look at current function Cannot maintain state across functions See also: docs/writinganllvmpass.html 24

25 Services provided by PassManager Optimization of pass execution: Process a function at a time instead of a pass at a time Example: If F, G, H are three functions in input pgm: FFFFGGGGHHHH not FGHFGHFGHFGH Process functions in parallel on an SMP (future work) Declarative dependency management: Automatically fulfill and manage analysis pass lifetimes Share analyses between passes when safe: e.g. DominatorSet live unless pass modifies CFG Avoid boilerplate for traversal of program See also: docs/writinganllvmpass.html 25

26 Pass Manager + Arg Promotion #1/2 Arg Promotion is a CallGraphSCCPass: Naturally operates bottom-up on the CallGraph Bubble pointers from callees out to callers 24: #include "llvm/callgraphsccpass.h" 47: struct SimpleArgPromotion : public CallGraphSCCPass { Arg Promotion requires AliasAnalysis info To prove safety of transformation Works with any alias analysis algorithm though 48: virtual void getanalysisusage(analysisusage &AU) const { AU.addRequired<AliasAnalysis>(); // Get aliases AU.addRequired<TargetData>(); // Get data layout CallGraphSCCPass::getAnalysisUsage(AU); // Get CallGraph 26

27 Pass Manager + Arg Promotion #2/2 Finally, implement runonscc (line 65): bool SimpleArgPromotion:: runonscc(const std::vector<callgraphnode*> &SCC) { bool Changed = false, LocalChange; do { // Iterate until we stop promoting from this SCC. LocalChange = false; // Attempt to promote arguments from all functions in this SCC. for (unsigned i = 0, e = SCC.size(); i!= e; ++i) LocalChange = PromoteArguments(SCC[i]); Changed = LocalChange; // Remember that we changed something. while (LocalChange); return Changed; // Passes return true if something changed. static int foo(int ***P) { return ***P; static int foo(int P_val_val_val) { return P_val_val_val; 27

28 Constant Propagation with DefUse Chains 28

29 LLVM Dataflow Analysis LLVM IR is in SSA form: use-def and def-use chains are always available All objects have user/use info, even functions Control Flow Graph is always available: Exposed as BasicBlock predecessor/successor lists Many generic graph algorithms usable with the CFG Higher-level info implemented as passes: Dominators, CallGraph, induction vars, aliasing, GVN, See also: docs/programmersmanual.html 29

30 Arg Promotion: safety check #1/4 #1: Function must be internal (aka static ) 88: if (!F!F->hasInternalLinkage()) return false; #2: Make sure address of F is not taken In LLVM, check that there are only direct calls using F 99: for (Value::use_iterator UI = F->use_begin(); UI!= F->use_end(); ++UI) { CallSite CS = CallSite::get(*UI); if (!CS.getInstruction()) // "Taking the address" of F. return false; #3: Check to see if any args are promotable: 114: for (unsigned i = 0; i!= PointerArgs.size(); ++i) if (!issafetopromoteargument(pointerargs[i])) PointerArgs.erase(PointerArgs.begin()+i); if (PointerArgs.empty()) return false; // no args promotable 30

31 Arg Promotion: safety check #2/4 #4: Argument pointer can only be loaded from: No stores through argument pointer allowed! // Loop over all uses of the argument (use-def chains). 138: for (Value::use_iterator UI = Arg->use_begin(); UI!= Arg->use_end(); ++UI) { // If the user is a load: if (LoadInst *LI = dyn_cast<loadinst>(*ui)) { // Don't modify volatile loads. if (LI->isVolatile()) return false; Loads.push_back(LI); else { return false; // Not a load. 31

32 Alias Analysis The AliasAnalysis class defines the interface that all alias analysis support Computed by the basicaa pass Can be changed Simple example int i; char C[2]; char A[10]; /*... */ for (i = 0; i!= 10; ++i) { C[0] = A[i]; /* One byte store */ C[1] = A[9-i]; /* One byte store */ 32

33 Arg Promotion: safety check #3/4 #5: Value of *P must not change in the BB We move load out to the caller, value cannot change! load P Modifie s *P? // Get AliasAnalysis implementation from the pass manager. 156: AliasAnalysis &AA = getanalysis<aliasanalysis>(); // Ensure *P is not modified from start of block to load 169: if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize)) return false; // Pointer is invalidated! See also: docs/aliasanalysis.html 33

34 Arg Promotion: safety check #4/4 #6: *P cannot change from Fn entry to BB Entry Modifie s *P? Modifie s *P? Entry load P load P 175: for (pred_iterator PI = pred_begin(bb), E = pred_end(bb); PI!= E; ++PI) // Loop over predecessors of BB. // Check each block from BB to entry (DF search on inverse graph). for (idf_iterator<basicblock*> I = idf_begin(*pi); I!= idf_end(*pi); ++I) // Might *P be modified in this basic block? if (AA.canBasicBlockModify(**I, Arg, LoadSize)) return false; 34

35 Arg Promotion: xform outline #1/4 #1: Make prototype with new arg types: #197 Basically just replaces int* with int in prototype #2: Create function with new prototype: 214: Function *NF = new Function(NFTy, F->getLinkage(), F->getName()); F->getParent()->getFunctionList().insert(F, NF); #3: Change all callers of F to call NF: // If there are uses of F, then calls to it remain. 221: while (!F->use_empty()) { // Get a caller of F. CallSite CS = CallSite::get(F->use_back()); 35

36 Arg Promotion: xform outline #2/4 #4: For each caller, add loads, determine args Loop over the args, inserting the loads in the caller 220: std::vector<value*> Args; 226: CallSite::arg_iterator AI = CS.arg_begin(); for (Function::aiterator I = F->abegin(); I!= F->aend(); ++I, ++AI) if (!ArgsToPromote.count(I)) // Unmodified argument. Args.push_back(*AI); else { // Insert the load before the call. LoadInst *LI = new LoadInst(*AI, (*AI)->getName()+".val", Call); // Insertion point Args.push_back(LI); 36

37 Arg Promotion: xform outline #3/4 #5: Replace the call site of F with call of NF // Create the call to NF with the adjusted arguments. 242: Instruction *New = new CallInst(NF, Args, "", Call); // If the return value of the old call was used, use the retval of the new call. if (!Call->use_empty()) Call->replaceAllUsesWith(New); // Finally, remove the old call from the program, reducing the use-count of F. Call->getParent()->getInstList().erase(Call); #6: Move code from old function to new Fn 259: NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 37

38 Arg Promotion: xform outline #4/4 #7: Change users of F s arguments to use NF s 264: for (Function::aiterator I = F->abegin(), I2 = NF->abegin(); I!= F->aend(); ++I, ++I2) if (!ArgsToPromote.count(I)) { // Not promoting this arg? I->replaceAllUsesWith(I2); // Use new arg, not old arg. else { while (!I->use_empty()) { // Only users can be loads. LoadInst *LI = cast<loadinst>(i->use_back()); LI->replaceAllUsesWith(I2); LI->getParent()->getInstList().erase(LI); #8: Delete old function: 286: F->getParent()->getFunctionList().erase(F); 38

39 Tutorial Overview Introduction to the running example LLVM C/C++ Compiler Overview High-level view of an example LLVM compiler The LLVM Virtual Instruction Set IR overview and type-system LLVM C++ IR and important API s Basics, PassManager, dataflow, ArgPromotion Alias Analysis in LLVM 39

40 Alias Analysis The AliasAnalysis class defines the interface that all alias analysis support Computed by the basicaa pass Can be changed Simple example int i; char C[2]; char A[10]; /*... */ for (i = 0; i!= 10; ++i) { C[0] = A[i]; /* One byte store */ C[1] = A[9-i]; /* One byte store */ 40

41 Project 1 Improve Pointer Analysis LLVM Study the precision of LLVM on some examples Suggest improvements by: Flow sensitivity Destructive updates 41

42 Project 2 Numeric analyzer in LLVM Integrate LLVM and Apron in a reasonable way Intera-procedural only Bonus interprocedural 42

43 Project 3 Shape Analysis in LLVM Integrate LLVM and TVLA in a reasonable way Intera-procedural only Generate TVP 43